Path following algorithms for $\ell_2$-regularized $M$-estimation with approximation guarantee

Authors: Yunzhang Zhu, Renxiong Liu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical results also corroborate our theoretical analysis.
Researcher Affiliation Collaboration Yunzhang Zhu Department of Statistics Ohio State University Columbus, OH 43015 zhu.219@osu.edu Renxiong Liu Statistics and Data Science team Nokia Bell Labs Murray Hill, NJ 07974 renxiong.liu@nokia-bell-labs.com
Pseudocode Yes Algorithm 1 A general path following algorithm. Input: ϵ > 0, C0 1/4, c1 1, c2 > 0, 0 < αmax 5 1 and tmax (0, ]. Output: grid points {tk}N k=1 and an approximated solution path θ(t). 1: Initialize: k = 1. 2: Compute α1 using (12). Starting from 0, iteratively calculate θ1 by minimizing ft1(θ) until (8) is satisfied for k = 1. 3: while (14) is not satisfied, do 4: Compute αk+1 using (13), update tk+1 = tk + αk+1. 5: Starting from θk, iteratively compute θk+1 by minimizing ftk+1(θ) until (8) is satisfied. 6: Update k = k + 1. 7: Interpolation: construct a solution path θ(t) through linear interpolation of {θk}N k=1 using (3).
Open Source Code No The paper mentions implementing methods in R using the Rcpp package and using LIBSVM, but does not state that its own source code for the methodology is openly available or provide a link.
Open Datasets Yes Example 3: Real data example. This example fits an ℓ2-regularized logistic regression using the a9a dataset from LIBSVM [Chang and Lin, 2011].
Dataset Splits No The paper mentions 'training data' and 'simulated datasets' but does not provide specific details on the training/validation/test splits (e.g., percentages, sample counts, or methodology).
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used for the experiments.
Software Dependencies No All methods considered in the numerical studies are implemented in R using Rcpp package [Eddelbuettel et al., 2011, Eddelbuettel, 2013]. While R and Rcpp are mentioned, specific version numbers for these software components are not provided.
Experiment Setup Yes Example 1: Ridge regression. This example considers ridge regression over a simulated dataset. In this case, the empirical loss function is Ln(θ) = Y Xθ 2 2/(2n), where X Rn p and Y Rn denote the design matrix and the response vector. In the simulation, the data (X, Y ) are generated from the usual linear regression model Y = Xθ + ϵ, where ϵ N(0, In n), θ = (1/ p, . . . , 1/ p) , and rows of X are IID samples from N(0, Ip p). Throughout this example, we consider n = 1000 and p = 500, 10000. ... Example 2: ℓ2-regularized logistic regression. ... We simulate the data (X, Y ) from a linear discriminant analysis (LDA) model. More specifically, we sample the components of Y independently from a Bernoulli distribution with P(Yi = +1) = 1/2; i = 1, 2, . . . , n. Conditioned on Yi, Xi s are then independently drawn from N(Yiµ, σ2Ip p), where µ Rp and σ2 > 0. ... Here we choose µ = (1/ p, . . . , 1/ p) and σ2 = 1 so that the Bayes risk is Φ( 1), which is approximately 15%. Similar to Example 1, two different problem dimension are considered: n = 1000 and p = 500, 10000. ... Throughout, we let tmax = 10.